Shadow AI: your employees use their own ChatGPT account, what should you do?

A salesperson pastes a proposal into their personal ChatGPT to rephrase it. An accountant drops an invoice export into it to get a summary. An HR manager has Claude proofread an interview report, on their phone. These three scenes are examples, but they happen every day: in France, 67% of employees who use AI rely on tools their company does not provide, often without telling it. In SMEs, the figure is 77% (OpinionWay for Cegid, September 2026).

This is not the employees’ fault: it shows they find a benefit in it. But that benefit comes with risks the company does not see, and it stays in accounts the company does not own.

What exactly is shadow AI?

Shadow AI is the use at work of AI tools the company has neither chosen nor governed: a personal ChatGPT, Claude or Gemini account, free or paid for by the employee, an app installed on their phone, a browser extension. The term comes from shadow IT. Microsoft also speaks of BYOAI, for Bring Your Own AI: 78% of AI users at work bring their own tools, 80% in small and medium-sized companies.

It is not occasional: 61% of French managers who use AI go through personal accounts at least once a week, and 38% every day (YouGov for Microsoft France, January 2026).

Why do your employees use their own account?

• The company provides nothing: only 39% of French employees have professional AI tools provided by their employer.
• The tool provided is less useful: generic or restricted, it knows neither your offers, nor your customers, nor the way you write. The personal account is already open on the phone.
• Nobody has said what is allowed: 63% of French employees say their company has set up no training on AI tools.
• They would rather not mention it: 52% of users are reluctant to admit they use AI for their most important tasks.

The seven real risks

1. Data that leaves, and may train the models

On a personal account, the employee sets the privacy options, not the company. On ChatGPT Free, Plus or Pro, the use of conversations to improve the models is switched on by default. On Claude Free, Pro and Max, since September 2025 it depends on a setting left to each user, and conversations are then kept for up to five years. The business plans of the same providers do not use your content to train their models by default. In 2023, Samsung engineers pasted internal source code into ChatGPT: the company banned these tools straight away.

2. The GDPR, with no contract and no register

As soon as an instruction given to the AI contains a name, an email address, a customer or employee file, it is processing of personal data, and your company remains responsible for it. With a personal account, no contract with the provider governs this processing (Article 28 of the GDPR), nothing appears in the register (Article 30), and you control neither where the data is processed, often in the United States, nor how long it is kept. If someone asks you what you did with their data (Articles 15 to 17), you cannot answer. The details: AI and the GDPR: what to check before deploying an agent.

3. Trade secrets and your confidentiality commitments weakened

Under French law, information is protected as a trade secret only if it is subject to "reasonable protective measures" (Commercial Code, Article L151-1). Prices, a strategy or a customer’s contract circulating in personal accounts weaken that argument. They may also breach the confidentiality clauses you have signed with your customers.

4. The company’s tools left unused

When the personal account remains more convenient, the official tool is not adopted: paid licences lie idle, and some employees pay for their subscription out of their own pocket. An MIT study showed it in 2025: in over 90% of the companies studied, employees regularly use personal AI tools, while only 40% have bought an official subscription. The gains exist, but nobody sees them, measures them or shares them with the rest of the team.

5. Know-how that leaves with the employee

The instructions that work, the reply templates, the history of exchanges: everything stays in the personal account. The day the employee leaves the company, that know-how leaves with them, and your data stays in their history, out of your reach.

6. Wrong answers sent out without review

A tool that does not know your rules invents with confidence: a price, a clause, a legal reference. Yet 31% of French employees already use AI to make important decisions in their work. Without a review rule, the error reaches the customer.

7. Security nobody manages, and evidence you cannot provide

No enforced authentication, no account closed when the employee leaves, no record. According to IBM, one organisation in five has already suffered a data breach due to shadow AI, and those where it is widespread paid an average of $670,000 more per breach. Gartner predicts that by 2030, more than 40% of organisations will suffer a security or compliance incident linked to unauthorised AI tools. And the day a customer, in a questionnaire or a tender, asks which AI tools you use and under which rules, you have nothing to show.

Why doesn’t banning work?

A ban without an alternative moves uses to phones and personal accounts: the gains remain, control disappears. Equipping, on the other hand, works. According to Netskope, between October 2024 and October 2025, the share of AI users going through personal accounts fell from 78% to 47%, while the share using company-provided accounts rose from 25% to 62%.

Your employees only leave their personal account for a tool that is more useful than it. That is where everything is decided.

The answer: assistants more useful than the personal account

Assistants set up for your company, within a well-tuned framework

A personal ChatGPT knows nothing about you. An assistant set up for a team knows your offers, prices, tone and procedures, and works within a framework you have defined. This framework, which specialists call the harness, comes down to five settings:

• business accounts, under contract with the provider, where your content is not used to train the models;
• privacy and access rights: who can see what, what is kept, what is closed when an employee leaves;
• the instructions given to the AI, written for each line of work: sales, customer service, administration;
• your reference documents (offers, prices, templates, procedures), kept on your side, outside the AI tool, which only receives a copy: you can change tools without losing anything;
• usage rules: what can be entrusted to the AI, what never goes into it, what a person reviews before it is sent.

Employees then get better answers from the company’s tool than from their own: they have no reason left to keep their personal account, and your data stays under contract. ChatGPT, Claude, Copilot or Mistral Vibe will do; Mistral Vibe, from Mistral AI, offers a European option.

A Company Brain, so that knowledge stays with you

To start with, an assistant’s reference documents are enough. When that knowledge has to serve the whole company, the Company Brain takes over: your company’s fundamentals (offers, prices, rules, priorities, tone) and the knowledge of each of your business domains, written down and kept up to date by a business domain lead. Your teams use it, your assistants and agents rely on it: everyone starts from the same rules. When an employee leaves, what they learned stays in the company. The Company Brain belongs to you, independent of any AI tool or model. To go further: Why your AI agents need a shared memory and shared rules.

The five-step method

1. Take stock, without sanctions

Meet the teams, look at the tools actually used, including those nobody has declared. The aim is to know, not to punish: those who already use AI are your best allies for what comes next.

2. Choose the tool and set it up

Start from what you already have (Microsoft 365, Google Workspace), your data constraints and the cost of licences. Then set up the business accounts, privacy and access rights.

3. Write a usage charter

One or two pages, no more: the tools allowed, the data that never goes into them (personal, confidential, strategic), who reviews what before it is sent, how to flag generated content.

4. Set up one assistant per team, then train

One assistant per line of work, fed with your reference documents and checked on real cases, then a hands-on session for each team. This is also what the AI Act requires under AI literacy: taking measures so that everyone masters the AI they use, in proportion to their use. See AI Act: what changes for SMEs and mid-sized companies, and when.

5. Measure and keep evidence

How many people use the tool each week, how much time it saves, and a register of the tools in use: that is what a client will ask for.

In a mid-sized company, a governance issue

The larger the company, the more uses multiply by department, with different tools. The subject becomes a question of consistency: the same rules for all, one way of answering customers, a register kept up to date by a named manager.

What we offer

• The AI Usage Audit takes stock of uses, including undeclared ones, and delivers the register, the technical gaps, a usage charter, a correction plan and a file of evidence, within five working days.
• The AI Assistant Setup puts in place the tool chosen with you, sets up one assistant per team on your reference documents, writes the usage charter and runs a hands-on session for each team, in 2 to 3 weeks.
• The Company Brain writes down your fundamentals and the knowledge of your business domains, for your teams as well as their assistants and agents.

Every offer is priced on a custom quote, after a free Scoping Meeting.

Sources

• OpinionWay for Cegid, "Le moment de vérité de l’IA en entreprise", 2,006 employees in France, Spain, Portugal and Germany, press release of 3 September 2026 (in French): in France, 67% of AI users rely on tools not provided by their company (77% in SMEs, 81% in mid-sized companies), 39% have professional tools, 63% have received no training, 31% use AI for important decisions.
• YouGov for Microsoft France, 657 managers and executives, January 2026, press release of 12 February 2026 (in French): 61% of AI users go through personal accounts at least once a week, 38% every day.
• Microsoft and LinkedIn, 2024 Work Trend Index, 31,000 people in 31 countries: 78% of AI users bring their own tools (80% in SMEs), 52% are reluctant to admit using AI for their most important tasks.
• MIT NANDA, "The GenAI Divide: State of AI in Business 2025", July 2025, as reported by Fortune.
• Netskope, Cloud and Threat Report 2026, data from October 2024 to October 2025.
• IBM, Cost of a Data Breach Report 2025, 600 organisations, breaches from March 2024 to February 2025.
• Gartner, November 2025 prediction on shadow AI incidents by 2030, as reported by Infosecurity Magazine.
• OpenAI, Data Controls FAQ; Anthropic, Updates to our consumer terms, August 2025.
• Bloomberg, Samsung bans staff’s AI use after spotting ChatGPT data leak, 2 May 2023.
• Regulation (EU) 2016/679 (GDPR), Articles 15 to 17, 28 and 30, official text; French Commercial Code, Article L151-1 (trade secrets); Regulation (EU) 2024/1689 on artificial intelligence, Article 4 (AI literacy), rewritten by Regulation (EU) 2026/1744, official text.